Transfer learning quiz

Question 1

What is transfer learning in machine learning?

Accepted Answer

Reusing a model trained on one task as a starting point for a related task

Answer

Moving data from one database to another before training

Answer

Training many small models and averaging their predictions

Answer

Compressing a model so it runs on mobile devices

Question 2

Why is transfer learning particularly valuable when you have a small dataset for your target task?

Accepted Answer

The pretrained model already learned general features, so little data is needed to adapt

Answer

Small datasets prevent overfitting automatically

Answer

Transfer learning removes the need for any labeled data at all

Answer

Small datasets always train faster than large ones regardless of method

Question 3

In image-based transfer learning, what kinds of features do the early layers of a pretrained CNN typically capture?

Accepted Answer

Edges, colors, and simple textures

Answer

Full object class labels

Answer

Specific dog vs cat decisions only

Answer

The final softmax probabilities

Question 4

What is the difference between 'feature extraction' and 'fine-tuning' in transfer learning?

Accepted Answer

Feature extraction freezes pretrained weights; fine-tuning updates them

Answer

Feature extraction needs more data than fine-tuning always

Answer

They are two names for the exact same process

Answer

Fine-tuning never changes any weights

Question 5

What does it mean to 'freeze' a layer in a neural network during transfer learning?

Accepted Answer

Its weights are kept fixed and not updated during training

Answer

Its weights are reset to random values each epoch

Answer

The layer is deleted from the network

Answer

The layer's outputs are ignored entirely

Question 6

ImageNet is often used as a source dataset for transfer learning. Why does this work well for many vision tasks?

Accepted Answer

It is large and diverse, so models learn general-purpose visual features

Answer

It contains exactly the images everyone needs for any task

Answer

It has no labels, which forces general learning

Answer

It is the only dataset that allows freezing layers

Question 7

When fine-tuning a pretrained model, why is it common to use a smaller learning rate than when training from scratch?

Accepted Answer

To gently adapt good pretrained weights without destroying them

Answer

Because pretrained models cannot use large learning rates technically

Answer

Because small learning rates always train faster

Answer

To force the model to forget the source task

Question 8

What is 'catastrophic forgetting' in the context of transfer learning and fine-tuning?

Accepted Answer

Losing previously learned knowledge while learning a new task

Answer

Forgetting to save a model checkpoint

Answer

Permanently freezing all layers by accident

Answer

A model crashing due to memory overflow

Question 9

Suppose your target task is very similar to the source task and you have a large dataset. What transfer learning strategy is generally recommended?

Accepted Answer

Fine-tune the whole network (or most of it)

Answer

Only use the model without any training

Answer

Freeze everything and train nothing

Answer

Train a completely new model from scratch and ignore pretraining

Question 10

You have a SMALL dataset and a target task that differs significantly from the source task. Why is this the hardest transfer learning scenario?

Accepted Answer

Few data points to fine-tune AND weak feature relevance from a different task

Answer

Large datasets always make transfer harder

Answer

Different tasks always share all features perfectly

Answer

The task is too easy so no learning is needed

Question 11

What is the role of the 'classifier head' (top layers) that you add when adapting a pretrained model to a new classification task?

Accepted Answer

It maps extracted features to the new task's specific classes

Answer

It freezes the pretrained layers automatically

Answer

It generates the raw input images

Answer

It deletes the feature extractor

Question 12

What is a 'pretrained model' and where can you typically obtain one?

Accepted Answer

A model with already-learned weights, available from hubs like Hugging Face or torchvision

Answer

A type of optimizer used only for fine-tuning

Answer

A dataset of unlabeled images

Answer

A model that has never been trained on any data

Question 13

In natural language processing, how does transfer learning typically work with models like BERT?

Accepted Answer

Pretrain on large text with self-supervision, then fine-tune on the target task

Answer

Translate all text to images first

Answer

Train fully from scratch on the target task only

Answer

Use the model with no pretraining at all

Question 14

What is 'self-supervised pretraining' and why is it important for transfer learning in NLP?

Accepted Answer

Generating labels automatically from the data, e.g. predicting masked words

Answer

Training only on small labeled datasets

Answer

Using a teacher network to label data

Answer

Manually labeling every sentence by hand

Question 15

What is 'domain adaptation', a specific type of transfer learning?

Accepted Answer

Same task but different data distribution between source and target

Answer

Completely different tasks with identical data

Answer

Changing the model architecture mid-training

Answer

Adapting hardware for faster training

Question 16

Why must input images usually be preprocessed (resized and normalized) to match what a pretrained model expects?

Accepted Answer

So inputs match the format the model was originally trained on

Answer

Normalization is optional and only affects speed

Answer

To make the dataset smaller

Answer

Because models reject any image that is not square

Question 17

What is 'gradual unfreezing' as a fine-tuning strategy?

Accepted Answer

Unfreezing layers progressively, starting from the top and moving deeper

Answer

Unfreezing all layers at once on the first epoch

Answer

Freezing layers only after training completes

Answer

Never unfreezing any layer

Question 18

How can transfer learning reduce the computational cost and training time of building a model?

Accepted Answer

You reuse learned features, so far fewer training iterations are needed

Answer

It removes the need for a GPU entirely

Answer

It increases the amount of data required

Answer

It makes inference slower but training faster

Question 19

In feature extraction transfer learning, what is the typical role of the pretrained network during training of the new classifier?

Accepted Answer

A fixed feature extractor whose outputs feed a trainable classifier

Answer

A data augmentation tool

Answer

A model that is fully retrained each step

Answer

A randomly reinitialized network

Question 20

A student fine-tunes a large pretrained model on only 50 images and gets 100% training accuracy but poor test accuracy. What is the most likely problem and a remedy?

Accepted Answer

Overfitting; remedy by freezing layers, augmentation, and regularization

Answer

Underfitting; remedy by removing all layers

Answer

The learning rate is too small; increase it greatly

Answer

The model is too small; make it much larger

Question 21

What does 'source domain' and 'target domain' mean in transfer learning terminology?

Accepted Answer

Source is the original training task/data; target is the new one

Answer

Both refer to the same dataset

Answer

Source is the GPU; target is the CPU

Answer

Source is the test set; target is the training set

Question 22

Why might transfer learning sometimes hurt performance, a phenomenon called 'negative transfer'?

Accepted Answer

When unrelated source knowledge misleads the model and lowers performance

Answer

When you freeze too few layers

Answer

When the dataset is too large

Answer

When the model trains too fast

Question 23

When adapting a pretrained classifier to a new task with a different number of classes, why must you replace the final output layer?

Accepted Answer

Its size matches the source classes, not your target's number of classes

Answer

Output layers cannot be frozen

Answer

The original layer has no activation function

Answer

It contains corrupted weights

Question 24

What is 'few-shot learning' and how does it relate to transfer learning?

Accepted Answer

Learning from very few examples, often enabled by pretrained representations

Answer

Training on millions of examples per class

Answer

A method to reduce model size

Answer

Learning that uses no examples at all

Question 25

In the Keras framework, what does setting 'include_top=False' do when loading a pretrained model like VGG16?

Accepted Answer

Loads the conv base but drops the original classification (top) layers

Answer

Freezes all layers automatically

Answer

Loads only the top classification layers

Answer

Loads the model with random weights

Question 26

How does data augmentation complement transfer learning when the target dataset is small?

Accepted Answer

It expands the small dataset with transformations, reducing overfitting

Answer

It deletes redundant images to shrink the dataset

Answer

It increases the number of model parameters

Answer

It replaces the need for a pretrained model

Question 27

Conceptually, why can the features learned by a deep network be 'reusable' across different tasks?

Accepted Answer

General low-level features (edges, shapes) are useful across many tasks

Answer

Only the output layer holds reusable knowledge

Answer

Features are random and meaningless

Answer

Every feature is unique to a single task

Question 28

What is the difference between 'inductive' and 'transductive' transfer learning at a basic level?

Accepted Answer

Inductive: different task with target labels; transductive: same task, different domain

Answer

They are identical concepts

Answer

Transductive always needs more labeled data than inductive

Answer

Inductive uses no neural networks; transductive does

Question 29

Why is it good practice to first train only the new head (with the base frozen) before unfreezing layers for fine-tuning?

Accepted Answer

The random head's large early gradients could corrupt pretrained weights

Answer

Heads must always stay frozen forever

Answer

Frozen bases cannot produce any output

Answer

It is required by all frameworks

Question 30

How does transfer learning relate to the idea of how humans learn new skills?

Accepted Answer

Reusing prior skills/knowledge to learn a related new skill faster

Answer

Learning each new skill completely independently

Answer

Forgetting old skills to learn new ones

Answer

Memorizing without understanding

Test yourself